What's this about?
LLMs increasingly mediate access to knowledge and computational assistance. The paper asks whether those capabilities reach vulnerable groups equally, and builds a multi-criteria benchmark for Braille comprehension in LLMs to find out.
Why is Braille a hard case for a language model?
Because Braille is not just an alphabet substitution. Contracted Braille compresses common letter groups and whole words into single cells, it has its own rules for numbers, capitalisation and mathematics, and the encoding differs by language and by grade. A model that has seen very little of it in training will produce confident, fluent, wrong output — which is worse than refusing, for a user who cannot easily check.
Why does this matter to you?
- Accessibility benchmarks are rare and this one is well-motivated: the groups who would benefit most are usually measured last.
- Multi-criteria is the right design — a single accuracy number would hide the failure modes that matter.
- Confident-and-wrong is the specific harm here, which is a general lesson about deploying models where users cannot verify.